Arc Pro B70 32GB

Intel · 32 GB GDDR6 · 608 GB/s · 225 W · MSRP $999

The Arc Pro B70 32GB is a Intel accelerator with 32 GB GDDR6 and about 608 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 39 curated models fit fully on this card; at FP16, about 18 fit. Top Q4 speeds: MiniCPM5-1B ≈ 358.2 tok/s (estimated, fits); Qwen3.5-2B ≈ 193.5 tok/s (estimated, fits); DeepSeek-R1-0528-Qwen3-8B-layer-mix-bpw-3.8-mlx ≈ 168.2 tok/s (estimated, fits); Qwen3.6-35B-A3B ≈ 129.0 tok/s (estimated, fits); Qwen3.5-35B-A3B ≈ 129.0 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

Fit checks

• Qwen3-8B — fits · ≈47.2 tok/s Q4 est.
• gemma-4-26B-A4B-it — fits · ≈101.8 tok/s Q4 est.
• Qwen2.5-7B-Instruct — fits · ≈50.9 tok/s Q4 est.
• Qwen3.5-9B — fits · ≈43.2 tok/s Q4 est.
• Llama-3.1-8B-Instruct — fits · ≈48.4 tok/s Q4 est.
• gpt-oss-20b — fits · ≈107.5 tok/s Q4 est.
• Qwen3.6-35B-A3B — fits · ≈129.0 tok/s Q4 est.
• Qwen3.5-35B-A3B — fits · ≈129.0 tok/s Q4 est.
• gemma-4-E2B-it — fits · ≈75.9 tok/s Q4 est.
• Qwen3.6-27B — fits · ≈13.9 tok/s Q4 est.
• Mistral-7B-Instruct-v0.3 — fits · ≈53.7 tok/s Q4 est.

All 69 model fit checks — see the linked Fits? column in the table above